void OrcBlockImpl::work(const InputItems &ins, const OutputItems &outs)
{
//calculate production/consumption params
size_t num_input_items, num_output_items;
if (_params.production_factor > 1.0)
{
num_output_items = std::min(size_t(ins.min()*_params.production_factor), outs.min());
num_input_items = size_t(num_output_items/_params.production_factor);
}
else
{
num_input_items = std::min(size_t(outs.min()/_params.production_factor), ins.min());
num_output_items = size_t(num_input_items*_params.production_factor);
}
//load the executor with source and dest buffers
size_t input_index = 0, output_index = 0;
for (size_t i = 0; i < ORC_N_VARIABLES; i++)
{
if (_orc_program->vars[i].vartype == ORC_VAR_TYPE_DEST)
{
void *dst = outs[output_index++].cast<void *>();
orc_executor_set_array(_orc_executor.get(), i, dst);
}
if (_orc_program->vars[i].vartype == ORC_VAR_TYPE_SRC)
{
void *src = const_cast<void *>(ins[input_index++].cast<const void *>());
orc_executor_set_array(_orc_executor.get(), i, src);
}
}
//execute the orc code;
orc_executor_set_n(_orc_executor.get(), size_t(_params.kernel_factor*num_input_items));
orc_executor_run(_orc_executor.get());
//produce consume fixed
this->consume(num_input_items-_params.consumption_offset);
this->produce(num_output_items);
}
void Multiply<std::complex<float> >::work(
const InputItems &ins, const OutputItems &outs
){
const size_t n_nums = std::min(ins.min(), outs.min());
std::complex<float> *out = outs[0].cast<std::complex<float> *>();
const std::complex<float> *in0 = ins[0].cast<const std::complex<float> *>();
for (size_t n = 1; n < ins.size(); n++)
{
const std::complex<float> *in = ins[n].cast<const std::complex<float> *>();
volk_32fc_x2_multiply_32fc(out, in0, in, n_nums * _vlen);
in0 = out; //for next input, we do output *= input
}
this->consume(n_nums);
this->produce(n_nums);
}
void Add<float>::work(
const InputItems &ins, const OutputItems &outs
){
const size_t n_nums = std::min(ins.min(), outs.min());
float *out = outs[0].cast<float *>();
const float *in0 = ins[0].cast<const float *>();
for (size_t n = 1; n < ins.size(); n++)
{
const float *in = ins[n].cast<const float *>();
volk_32f_x2_add_32f(out, in0, in, n_nums * _vlen);
in0 = out; //for next input, we do output += input
}
this->consume(n_nums);
this->produce(n_nums);
}
void Multiply<type>::work(
const InputItems &ins, const OutputItems &outs
){
const size_t n_nums = std::min(ins.min(), outs.min());
type *out = outs[0].cast<type *>();
const type *in0 = ins[0].cast<const type *>();
for (size_t n = 1; n < ins.size(); n++)
{
const type *in = ins[n].cast<const type *>();
for (size_t i = 0; i < n_nums * _vlen; i++)
{
out[i] = in0[i] * in[i];
}
in0 = out; //for next input, we do output *= input
}
this->consume(n_nums);
this->produce(n_nums);
}